Long-form devlogs

Load video-craft (pipeline, tools, QA gate) and video-youtube (title, thumbnail, upload). This skill covers the episode itself.

Is it worth making yet?

• Devlogs are mid-funnel: they deepen people who already know the game. Reported conversion is roughly 0.07-1.3% of views to wishlists, and the strongest devlog results came from years of steady uploads [PRAC]. Zukowski's survey ranks festivals, TikTok, streamers and Reddit above YouTube devlogs, and Jonas Tyroller says his own channel made no noticeable dent for Thronefall. The evidence conflicts, so run a small, player-facing series at a pace you can hold and judge it on weekly wishlists once the page exists.
• Start long-form only when there is enough approved gameplay for an episode. Until then, devlog topics become Shorts.
• Voice: your own voice, captions only, or an AI voice. If AI, use one approved voice and settings, and disclose it (YouTube AI use: Yes). An AI-sounding read costs trust with a dev audience, so the script and the take matter more: short lines, several takes, chosen by ear (elevenlabs skill, voice section).

One episode = one topic

• A topic is a player-facing moment, not an engine topic: "This hero is hiding in a diner", not "Godot shaders". Good topics: a character or enemy reveal, a system that was broken and is fixed, a visual overhaul with before/after on approved art, a playtest invite.
• Sprint logs and bug-fix lists underperform. Tangents become notes for the next episode.
• Keep a running devlog-notes.md: one line and a clip path whenever something visible changes. If battles are deterministic, fixed seeds plus git history let you re-shoot any past state later instead of recording while working.

Structure (8-12 min)

• YouTube's Intro metric is the share still watching at 0:30, measured against what the title and thumbnail promised. Pay the promise inside 30 s.
• Chapters: 00:00 first, at least 3, each ≥ 10 s, named by question ("Why the fight felt dead").

Script

• Write the body first and the cold open last, from the real payoff.
• About 120-150 spoken words a minute: 600-1,200 words for 5-10 min. Read it aloud; cut every sentence you stumble on.
• Show, then tell: wherever gameplay can replace a sentence, it does. No talking over a static screen.
• Plain, first person, no trailer voice, no invented facts or anecdotes. Run a humanizer pass for the script's language (e.g. https://github.com/blader/humanizer), and get the human's approval on the script before any paid rendering.
• Captions-only episodes: one caption per thought, ≤ 2 lines, in sync with the picture; a captions-only video has to move faster because nothing carries the gaps.

Retention

• New visual information every 3-6 s; change the shot when the sentence changes topic.
• A pattern interrupt every 30-60 s: an angle change, B-roll, a diagram, a music change, a speed ramp, a held silence.
• Each section ends on an open question the next one answers.
• Music under speech stays low (start at −20 dB under the voice and measure, video-craft tools.md "Measuring sound") or drops out during dense explanation.
• Narration gaps: at most 0.5 s between lines, keeping the breath at a sentence end. With TTS, take word times from the provider's timestamp response so cuts and captions never clip a word.
• One 1.5× watch-through before export: cut anything you'd skip. Write every fix in notes.md during that watch and fix them in one round.
• Build each chapter as its own segment and concat them, so a fix in one chapter re-renders only that chapter.
• After publishing, retention spikes are the next Shorts and the next thumbnail ideas.

Every episode also makes

• 1-2 vertical Shorts (video-short): one moment, the hook in the first second, the game named on screen, the episode set as the Short's Related video.
• An X clip and, where it fits a sub's rules, a Reddit post, each cut for its channel.
• Later, the same topic as a Steam event post so the store page shows activity.

Checklist (on top of the video-craft QA gate)

• The title and thumbnail promise is on screen by 0:30, and the game is named by 0:30.
• One topic; every section serves it.
• Chapters valid; the ask comes before the outro.
• Script approved; narration in the approved voice and settings; YouTube AI use set to Yes if the voice is synthetic.
• Music and SFX are your own or licensed on file; loudness −14 LUFS.
• Privacy pass on any screen capture (editor, terminal, notes, mail): names, e-mails, paths, keys and tokens blurred (video-craft tools.md blur recipe), checked frame by frame where they move.
• An SRT is ready, plus translated titles and descriptions for your target languages (video-youtube).
• The Shorts cut from it are made.

Reading the result

Read at 48 h and 7 d: Intro (% at 0:30), CTR by traffic source, average view duration, dips and spikes. Judge the series after 5-10 episodes on weekly wishlist change (0-40 a week is weak, 200+ is strong by Zukowski's benchmark), not on views. If an episode's Intro is low, the cold open didn't pay the title; fix that first.